AgentRadio Messaging Nearly Doubles Coding-Agent Task Score

AgentRadio lets coding agents message each other mid-task, and four coordinated agents resolved 62.1% of codebase questions versus 32.3% for one agent working alone, per a July 30 arXiv paper. That's a nearly 30-point jump in benchmark accuracy. It comes from a coordination layer teams could add to agents they already run, not a newer model.

The system adds three primitives: threads, messages, and waiting for a mention. The wait runs as a background task, so an agent keeps working in the foreground instead of stopping to check in.

The benchmark is SWE-Atlas QnA, a set of long-horizon questions about production codebases. The four-agent group split the work using a five-phase division-of-labor protocol, and the code is released under Coral-Protocol.

That 62.1% score also beats a single Claude Code agent running the newer Opus 4.8 model, which answers 57.2% of the same questions.

The accuracy gap between solo and coordinated agents widens as tasks get harder, per the paper. The authors read that as mid-course correction, agents catching and fixing each other's wrong turns, not four agents splitting a workload.

A related paper, SkillRise, takes a different route to the same goal. It has a single agent alternate between solving tasks and rewriting its own skill document. No teammates required, one policy improving itself instead of four agents talking to each other.